{"id":"https://openalex.org/W7166829037","doi":"https://doi.org/10.18653/v1/2026.findings-acl.880","title":"Reinforcement Learning with Semantic Rewards Enables Low-Resource Language Expansion without Alignment Tax","display_name":"Reinforcement Learning with Semantic Rewards Enables Low-Resource Language Expansion without Alignment Tax","publication_year":2026,"publication_date":"2026-01-01","ids":{"openalex":"https://openalex.org/W7166829037","doi":"https://doi.org/10.18653/v1/2026.findings-acl.880"},"language":null,"primary_location":{"id":"doi:10.18653/v1/2026.findings-acl.880","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.findings-acl.880","pdf_url":"https://aclanthology.org/2026.findings-acl.880.pdf","source":null,"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Findings of the Association for Computational Linguistics: ACL 2026","raw_type":"proceedings-article"},"type":"conference-paper","indexed_in":["crossref"],"open_access":{"is_oa":true,"oa_status":"gold","oa_url":"https://aclanthology.org/2026.findings-acl.880.pdf","any_repository_has_fulltext":null},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5121319865","display_name":"Zeli Su","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zeli Su","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5139839329","display_name":"Ziyin Zhang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Ziyin Zhang","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5139737763","display_name":"Zhou Liu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zhou Liu","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5136214581","display_name":"Xuexian Song","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Xuexian Song","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5136219478","display_name":"Zhankai Xu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zhankai Xu","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5139760979","display_name":"Longfei Zheng","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Longfei Zheng","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5139717557","display_name":"Xiaolu Zhang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Xiaolu Zhang","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5139779677","display_name":"Rong Fu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Rong Fu","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5139726427","display_name":"Guixian Xu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Guixian Xu","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5139787093","display_name":"Wentao Zhang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wentao Zhang","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]}],"institutions":[],"countries_distinct_count":0,"institutions_distinct_count":0,"corresponding_author_ids":[],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":0.0,"has_fulltext":true,"cited_by_count":0,"citation_normalized_percentile":{"value":0.81867752,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":"17772","last_page":"17786"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10181","display_name":"Natural Language Processing Techniques","score":0.374099999666214,"subfield":{"id":"https://openalex.org/subfields/1702","display_name":"Artificial Intelligence"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},"topics":[{"id":"https://openalex.org/T10181","display_name":"Natural Language Processing Techniques","score":0.374099999666214,"subfield":{"id":"https://openalex.org/subfields/1702","display_name":"Artificial Intelligence"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},{"id":"https://openalex.org/T10028","display_name":"Topic Modeling","score":0.09939999878406525,"subfield":{"id":"https://openalex.org/subfields/1702","display_name":"Artificial Intelligence"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},{"id":"https://openalex.org/T12072","display_name":"Machine Learning and Algorithms","score":0.07739999890327454,"subfield":{"id":"https://openalex.org/subfields/1702","display_name":"Artificial Intelligence"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}}],"keywords":[{"id":"https://openalex.org/keywords/reinforcement-learning","display_name":"Reinforcement learning","score":0.4903999865055084},{"id":"https://openalex.org/keywords/semantics","display_name":"Semantics (computer science)","score":0.3352999985218048},{"id":"https://openalex.org/keywords/action","display_name":"Action (physics)","score":0.2994999885559082},{"id":"https://openalex.org/keywords/class","display_name":"Class (philosophy)","score":0.2937000095844269},{"id":"https://openalex.org/keywords/variety","display_name":"Variety (cybernetics)","score":0.28360000252723694},{"id":"https://openalex.org/keywords/reinforcement","display_name":"Reinforcement","score":0.27869999408721924}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6085000038146973},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.527400016784668},{"id":"https://openalex.org/C97541855","wikidata":"https://www.wikidata.org/wiki/Q830687","display_name":"Reinforcement learning","level":2,"score":0.4903999865055084},{"id":"https://openalex.org/C204321447","wikidata":"https://www.wikidata.org/wiki/Q30642","display_name":"Natural language processing","level":1,"score":0.34360000491142273},{"id":"https://openalex.org/C184337299","wikidata":"https://www.wikidata.org/wiki/Q1437428","display_name":"Semantics (computer science)","level":2,"score":0.3352999985218048},{"id":"https://openalex.org/C2780791683","wikidata":"https://www.wikidata.org/wiki/Q846785","display_name":"Action (physics)","level":2,"score":0.2994999885559082},{"id":"https://openalex.org/C2777212361","wikidata":"https://www.wikidata.org/wiki/Q5127848","display_name":"Class (philosophy)","level":2,"score":0.2937000095844269},{"id":"https://openalex.org/C136197465","wikidata":"https://www.wikidata.org/wiki/Q1729295","display_name":"Variety (cybernetics)","level":2,"score":0.28360000252723694},{"id":"https://openalex.org/C67203356","wikidata":"https://www.wikidata.org/wiki/Q1321905","display_name":"Reinforcement","level":2,"score":0.27869999408721924},{"id":"https://openalex.org/C2775924081","wikidata":"https://www.wikidata.org/wiki/Q55608371","display_name":"Control (management)","level":2,"score":0.275299996137619},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.2678000032901764},{"id":"https://openalex.org/C165064840","wikidata":"https://www.wikidata.org/wiki/Q1321061","display_name":"Matching (statistics)","level":2,"score":0.26440000534057617},{"id":"https://openalex.org/C2776401178","wikidata":"https://www.wikidata.org/wiki/Q12050496","display_name":"Feature (linguistics)","level":2,"score":0.2574000060558319}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.18653/v1/2026.findings-acl.880","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.findings-acl.880","pdf_url":"https://aclanthology.org/2026.findings-acl.880.pdf","source":null,"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Findings of the Association for Computational Linguistics: ACL 2026","raw_type":"proceedings-article"}],"best_oa_location":{"id":"doi:10.18653/v1/2026.findings-acl.880","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.findings-acl.880","pdf_url":"https://aclanthology.org/2026.findings-acl.880.pdf","source":null,"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Findings of the Association for Computational Linguistics: ACL 2026","raw_type":"proceedings-article"},"sustainable_development_goals":[],"awards":[],"funders":[],"has_content":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7166829037.pdf","grobid_xml":"https://content.openalex.org/works/W7166829037.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Extending":[0],"large":[1],"language":[2,17,178],"models":[3],"(LLMs)":[4],"to":[5],"low-resource":[6,113,177],"languages":[7],"often":[8],"incurs":[9],"an":[10],"\"alignment":[11],"tax\":":[12],"improvements":[13],"in":[14,25,140],"the":[15,20,34,68],"target":[16],"come":[18],"at":[19],"cost":[21],"of":[22,36],"catastrophic":[23],"forgetting":[24],"general":[26,121],"capabilities.We":[27],"argue":[28],"that":[29,90,109,148,162],"this":[30,52],"trade-off":[31],"arises":[32],"from":[33],"rigidity":[35],"supervised":[37],"fine-tuning":[38],"(SFT),":[39],"which":[40],"enforces":[41],"token-level":[42],"surface":[43,130],"imitation":[44],"on":[45,100],"narrow":[46],"and":[47,104,138,143,153,171],"biased":[48],"data":[49],"distributions.To":[50],"address":[51],"limitation,":[53],"we":[54],"propose":[55],"a":[56,169],"semantic-space":[57],"alignment":[58,118],"paradigm":[59],"powered":[60],"by":[61],"Group":[62],"Relative":[63],"Policy":[64],"Optimization":[65],"(GRPO),":[66],"where":[67],"model":[69],"is":[70],"optimized":[71],"using":[72],"embedding-level":[73],"semantic":[74,132,136,166],"rewards":[75,167],"rather":[76],"than":[77,125],"likelihood":[78],"maximization.This":[79],"objective":[80],"encourages":[81],"meaning":[82],"preservation":[83],"through":[84],"flexible":[85],"realizations,":[86],"enabling":[87],"controlled":[88],"updates":[89],"reduce":[91],"destructive":[92],"interference":[93],"with":[94,165],"pretrained":[95],"knowledge.We":[96],"evaluate":[97],"our":[98,110,159],"approach":[99],"Tibetan-Chinese":[101],"machine":[102],"translation":[103],"Tibetan":[105],"headline":[106],"generation.Experiments":[107],"show":[108],"method":[111],"acquires":[112],"capabilities":[114],"while":[115],"markedly":[116],"mitigating":[117],"tax,":[119],"preserving":[120],"competence":[122],"more":[123,151,172],"effectively":[124],"SFT.Despite":[126],"producing":[127],"less":[128],"rigid":[129],"overlap,":[131],"RL":[133],"yields":[134],"higher":[135],"quality":[137],"preference":[139],"open-ended":[141],"generation,":[142],"few-shot":[144],"transfer":[145],"results":[146],"indicate":[147],"it":[149],"learns":[150],"transferable":[152],"robust":[154],"representations":[155],"under":[156],"limited":[157],"supervision.Overall,":[158],"study":[160],"demonstrates":[161],"reinforcement":[163],"learning":[164],"provides":[168],"safer":[170],"reliable":[173],"pathway":[174],"for":[175],"inclusive":[176],"expansion.":[179]},"counts_by_year":[],"updated_date":"2026-07-29T14:22:42.915294","created_date":"2026-07-02T00:00:00"}